Classification of infrared spectra of organophosphorus compounds with artificial neural networks

We examine the use of artificial neural networks to classify infrared spectra of organophosphorus pesticides and chemically related compounds. The spectra used were contributed from commercial libraries, government agencies, and government contractors and include spectra of pesticides, industrial precursors, hydrolysis products and other organophosphorus compounds. The data were pretreated to reduce artifacts arising from the variety of collection sources. The treated spectra were divided into spectral "bins" of equal frequency width and transduced into data vectors whose elements consisted of the average absorbance value of the corresponding spectral bin, The spectral data vectors served as inputs to neural networks examined as spectral classifiers. Conventional feed-forward neural networks were constructed using the spectral data vectors as inputs and indicating the chemical class in the output nodes. Several training methods were compared for efficiency in setting the network weights and biases, from the standpoint of time needed to train the network and from the error frequencies obtained Network architecture was also examined in order to obtain the optimum number of nodes to place in the hidden layers of the networks. Radial basis function networks were also generated from spectral input data and used to classify the spectra. The final classification accuracies and training efficiencies from these networks were compared with those from the feed forward networks. The radial basis function networks could be trained rapidly and produced numerically stable networks, but the network generation technique tended to produce large numbers of hidden nodes. The radial basis function networks gave the best classification accuracy obtained with neural network classifiers.